ggml : scratch that - vmlaq_n_f32 is always better

Had a background process that was messing with the timings

.github/workflows/build.yml

@@ -8,6 +8,8 @@ on:
         required: true
         type: boolean
   push:
+    branches:
+      - master
     paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
   pull_request:
     types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
@@ -18,6 +20,8 @@ env:
 
 jobs:
   ubuntu-latest-make:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     steps:
@@ -37,6 +41,8 @@ jobs:
           make
 
   ubuntu-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     steps:
@@ -65,6 +71,8 @@ jobs:
           ctest --verbose
 
   ubuntu-latest-cmake-sanitizer:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     continue-on-error: true
@@ -101,6 +109,8 @@ jobs:
           ctest --verbose
 
   macOS-latest-make:
+    if: github.event.pull_request.draft == false
+
     runs-on: macos-latest
 
     steps:
@@ -119,6 +129,8 @@ jobs:
           make
 
   macOS-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: macOS-latest
 
     steps:
@@ -146,6 +158,8 @@ jobs:
           ctest --verbose
 
   windows-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: windows-latest
 
     strategy:

.github/workflows/docker.yml

@@ -18,6 +18,8 @@ on:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
     env:
       COMMIT_SHA: ${{ github.sha }}

.gitignore

@@ -24,6 +24,7 @@ models/*
 /perplexity
 /embedding
 /benchmark-q4_0-matmult
+/vdot
 /Pipfile
 
 arm_neon.h

CMakeLists.txt

@@ -55,6 +55,8 @@ option(LLAMA_SANITIZE_UNDEFINED     "llama: enable undefined sanitizer"
 option(LLAMA_AVX                    "llama: enable AVX"                                     ON)
 option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
 option(LLAMA_AVX512                 "llama: enable AVX512"                                  OFF)
+option(LLAMA_AVX512_VBMI            "llama: enable AVX512-VBMI"                             OFF)
+option(LLAMA_AVX512_VNNI            "llama: enable AVX512-VNNI"                             OFF)
 option(LLAMA_FMA                    "llama: enable FMA"                                     ON)
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
@@ -220,6 +222,16 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     if (MSVC)
         if (LLAMA_AVX512)
             add_compile_options(/arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (LLAMA_AVX512_VBMI)
+                add_compile_definitions(__AVX512VBMI__)
+            endif()
+            if (LLAMA_AVX512_VNNI)
+                add_compile_definitions(__AVX512VNNI__)
+            endif()
         elseif (LLAMA_AVX2)
             add_compile_options(/arch:AVX2)
         elseif (LLAMA_AVX)
@@ -240,9 +252,13 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
         endif()
         if (LLAMA_AVX512)
             add_compile_options(-mavx512f)
-            # add_compile_options(-mavx512cd)
-            # add_compile_options(-mavx512dq)
-            # add_compile_options(-mavx512bw)
+            add_compile_options(-mavx512bw)
+        endif()
+        if (LLAMA_AVX512_VBMI)
+            add_compile_options(-mavx512vbmi)
+        endif()
+        if (LLAMA_AVX512_VNNI)
+            add_compile_options(-mavx512vnni)
         endif()
     endif()
 else()
@@ -289,4 +305,5 @@ endif ()
 
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
+    add_subdirectory(pocs)
 endif()

Makefile

@@ -133,7 +133,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize quantize-stats perplexity embedding
+default: main quantize quantize-stats perplexity embedding vdot
 
 #
 # Build library
@@ -169,6 +169,9 @@ perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
 embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
+vdot: pocs/vdot/vdot.cpp ggml.o
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+
 libllama.so: llama.o ggml.o
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 

README.md

@@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
+- [Added LoRA support](https://github.com/ggerganov/llama.cpp/pull/820)
 - [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915)
 - [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784)
 
@@ -50,6 +51,7 @@ New features will probably be added mostly through community contributions.
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 
 **UI:**
 

convert-lora-to-ggml.py

@@ -0,0 +1,124 @@
+import json
+import os
+import re
+import struct
+import sys
+from typing import Any, Dict, Sequence, TextIO
+
+import torch
+
+from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType
+
+HF_SUBLAYER_TO_GGML = {
+    "self_attn.q_proj": "attention.wq",
+    "self_attn.k_proj": "attention.wk",
+    "self_attn.v_proj": "attention.wv",
+    "self_attn.o_proj": "attention.wo",
+    "mlp.gate_proj": "feed_forward.w1",
+    "mlp.down_proj": "feed_forward.w2",
+    "mlp.up_proj": "feed_forward.w3",
+    "input_layernorm": "attention_norm",
+    "post_attention_layernorm": "ffn_norm",
+    # "norm": "norm",
+    # "embed_tokens": "tok_embeddings",
+    # "lm_head": "output",
+}
+
+
+def translate_tensor_name(t: str) -> str:
+    match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
+    if match:
+        nn = match.group(1)
+        sub_layer = match.group(2)
+        lora_type = match.group(3)
+
+        sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
+        if sub_layer_renamed is None:
+            print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
+            sys.exit(1)
+
+        output_string = (
+            f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+        )
+        return output_string
+    else:
+        print(f"Error: unrecognized tensor {t}")
+        sys.exit(1)
+
+
+def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+    fout.write(b"ggla"[::-1])  # magic (ggml lora)
+    fout.write(struct.pack("i", 1))  # file version
+    fout.write(struct.pack("ii", params["r"], params["lora_alpha"]))
+
+
+def write_tensor_header(
+    self, name: str, shape: Sequence[int], data_type: DataType
+) -> None:
+    sname = name.encode("utf-8")
+    fout.write(
+        struct.pack(
+            "iii",
+            len(shape),
+            len(sname),
+            DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]],
+        )
+    )
+    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
+    fout.write(sname)
+    fout.seek((fout.tell() + 31) & -32)
+
+
+if len(sys.argv) != 2:
+    print(f"Usage: python {sys.argv[0]} <path>")
+    print(
+        "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
+    )
+    sys.exit(1)
+
+input_json = os.path.join(sys.argv[1], "adapter_config.json")
+input_model = os.path.join(sys.argv[1], "adapter_model.bin")
+output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+
+model = torch.load(input_model, map_location="cpu")
+
+with open(input_json, "r") as f:
+    params = json.load(f)
+
+if params["peft_type"] != "LORA":
+    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+    sys.exit(1)
+
+if params["fan_in_fan_out"] == True:
+    print("Error: param fan_in_fan_out is not supported")
+    sys.exit(1)
+
+if params["bias"] is not None and params["bias"] != "none":
+    print("Error: param bias is not supported")
+    sys.exit(1)
+
+# TODO: these seem to be layers that have been trained but without lora.
+# doesn't seem widely used but eventually should be supported
+if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
+    print("Error: param modules_to_save is not supported")
+    sys.exit(1)
+
+with open(output_path, "wb") as fout:
+    fout.truncate()
+
+    write_file_header(fout, params)
+    for k, v in model.items():
+        if k.endswith("lora_A.weight"):
+            if v.dtype != torch.float16 and v.dtype != torch.float32:
+                v = v.float()
+            v = v.T
+        else:
+            v = v.float()
+
+        t = v.numpy()
+        tname = translate_tensor_name(k)
+        print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+        write_tensor_header(fout, tname, t.shape, t.dtype)
+        t.tofile(fout)
+
+print(f"Converted {input_json} and {input_model} to {output_path}")

convert.py

@@ -735,7 +735,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
     header: Dict[str, Dict[str, Any]] = json.loads(fp.read(header_size))
     # Use mmap for the actual data to avoid race conditions with the file offset.
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
-    byte_buf = mapped[fp.tell():]
+    byte_buf = mapped[8 + header_size:]
 
     def convert(info: Dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
@@ -761,7 +761,7 @@ def must_read(fp: IO[bytes], length: int) -> bytes:
     return ret
 
 
-def lazy_load_ggml_file(fp: IO[bytes], path: Path) -> ModelPlus:
+def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
     magic = must_read(fp, 4)[::-1]
     if magic in (b'ggmf', b'ggjt'):
         version, = struct.unpack("i", must_read(fp, 4))
@@ -795,7 +795,9 @@ def lazy_load_ggml_file(fp: IO[bytes], path: Path) -> ModelPlus:
 
     model: LazyModel = {}
     # Use mmap for the actual data to avoid race conditions with the file offset.
+    off = fp.raw.tell()
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
+    fp.raw.seek(off) # needed on Windows
 
     def read_tensor() -> None:  # this is a function so that variables captured in `load` don't change
         shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
@@ -949,8 +951,9 @@ class OutputFile:
 
         ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            size = ' x '.join(map(str, lazy_tensor.shape))
-            print(f"[{i+1}/{len(model)}] Writing tensor {name}, size {size}...")
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
             of.write_tensor_header(name, lazy_tensor.shape, lazy_tensor.data_type)
             ndarray.tofile(of.fout)
         of.fout.close()
@@ -1082,6 +1085,7 @@ def default_outfile(model_paths: List[Path], params: Params) -> Path:
     namestr = {
         GGMLFileType.AllF32: "f32",
         GGMLFileType.MostlyF16: "f16",
+        GGMLFileType.MostlyQ4_0: "q4_0",
         GGMLFileType.MostlyQ4_1: "q4_1",
         GGMLFileType.PerLayerIsQ4_1: "q4_1",
     }[params.file_type]
@@ -1105,7 +1109,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    parser.add_argument("--outtype", choices=["f32", "f16", "q4_1"], help="output format (default: based on input)")
+    parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)")
     parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
     parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")

examples/common.cpp

@@ -139,6 +139,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter = argv[i];
+            params.use_mmap = false;
+        } else if (arg == "--lora-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
         } else if (arg == "-i" || arg == "--interactive") {
             params.interactive = true;
         } else if (arg == "--embedding") {
@@ -242,6 +255,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     }
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "\n");

examples/common.h

@@ -31,11 +31,12 @@ struct gpt_params {
 
     std::string model  = "models/lamma-7B/ggml-model.bin"; // model path
     std::string prompt = "";
-    std::string input_prefix = ""; // string to prefix user inputs with
-
-
+    std::string input_prefix = "";       // string to prefix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
+    std::string lora_adapter = "";  // lora adapter path
+    std::string lora_base = "";     // base model path for the lora adapter
+
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

examples/embedding/embedding.cpp

@@ -1,6 +1,8 @@
 #include "common.h"
 #include "llama.h"
 
+#include <ctime>
+
 int main(int argc, char ** argv) {
     gpt_params params;
     params.model = "models/llama-7B/ggml-model.bin";

examples/main/main.cpp

@@ -11,6 +11,7 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
+#include <ctime>
 #include <fstream>
 #include <iostream>
 #include <string>
@@ -113,6 +114,17 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if (!params.lora_adapter.empty()) {
+        int err = llama_apply_lora_from_file(ctx,
+                                             params.lora_adapter.c_str(),
+                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             params.n_threads);
+        if (err != 0) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+            return 1;
+        }
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");

examples/perplexity/perplexity.cpp

@@ -2,6 +2,7 @@
 #include "llama.h"
 
 #include <cmath>
+#include <ctime>
 
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
@@ -133,6 +134,17 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if (!params.lora_adapter.empty()) {
+        int err = llama_apply_lora_from_file(ctx,
+                                             params.lora_adapter.c_str(),
+                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             params.n_threads);
+        if (err != 0) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+            return 1;
+        }
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");

examples/quantize-stats/quantize-stats.cpp

@@ -221,7 +221,7 @@ int main(int argc, char ** argv) {
                 break;
             }
             int j;
-            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) i)) != 0; j++) {
+            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) {
                 // find match
             }
             if (j < GGML_TYPE_COUNT) {

ggml.c

@@ -1420,6 +1420,34 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in
 #endif
 }
 
+static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
+    [GGML_TYPE_Q4_0] = {
+        .dequantize_row_q         = dequantize_row_q4_0,
+        .quantize_row_q           = quantize_row_q4_0,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference,
+        .quantize_row_q_dot       = quantize_row_q8_0,
+        .vec_dot_q                = ggml_vec_dot_q4_0_q8_0,
+    },
+    [GGML_TYPE_Q4_1] = {
+        .dequantize_row_q         = dequantize_row_q4_1,
+        .quantize_row_q           = quantize_row_q4_1,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference,
+        .quantize_row_q_dot       = quantize_row_q4_1,
+        .vec_dot_q                = ggml_vec_dot_q4_1,
+    },
+    // TODO: GGML_TYPE_Q8_0
+};
+
+// For internal test use
+quantize_fns_t ggml_internal_get_quantize_fn(size_t i) {
+    GGML_ASSERT(i < GGML_TYPE_COUNT);
+    return quantize_fns[i];
+}
+
+
 //
 // simd mappings
 //
@@ -1977,33 +2005,187 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float
 }
 
 #if __AVX512F__ && QK4_0 == 32
-static inline __m512 dot_q4_0_oneblock_avx512(
+static inline __m512i bytes_from_q4_0_twoblocks_avx512( const __m512i blocks ) {
+    // The 64 bytes of `blocks` contain two consecutive Q4_0 blocks loaded from memory:
+    // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+    // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32|
+    // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+    // |                                                                        :. =_ () [] <> () Zz Yy|
+    // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+    // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00|
+    // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+    // |Xx Ww Vv Uu Tt Ss Rr Qq             Pp Oo Nn Mm Ll Kk Jj Ii Hh Gg Ff Ee Dd Cc Bb Aa            |
+    // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+    //
+    // Bytes 04..19 (block #0) and 24..39 (block #1) both contain 32 nibbles (4-bit unsigned integers).
+    // We have exactly 64 nibbles, so we want to place each nibble into a separate byte.
+    // Bytes 00..03 and 20..23 contain scales, which are irrelevant to this function.
+    // Bytes 40..63 are masked when loading the data, so they are zeroed out.
+#ifdef __AVX512VBMI__
+    const __m512i byte_perm = _mm512_set_epi8(
+        39, 38, 39, 38, 37, 36, 37, 36, 35, 34, 35, 34, 33, 32, 33, 32,
+        31, 30, 31, 30, 29, 28, 29, 28, 27, 26, 27, 26, 25, 24, 25, 24,
+        19, 18, 19, 18, 17, 16, 17, 16, 15, 14, 15, 14, 13, 12, 13, 12,
+        11, 10, 11, 10,  9,  8,  9,  8,  7,  6,  7,  6,  5,  4,  5,  4
+    );
+    const __m512i permuted = _mm512_permutexvar_epi8( byte_perm, blocks );
+    // After applying VPERMB, `permuted` looks like this:
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |:. =_ :. =_ () [] () [] <> () <> () Zz Yy Zz Yy Xx Ww Xx Ww Vv Uu Vv Uu Tt Ss Tt Ss Rr Qq Rr Qq|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |Pp Oo Pp Oo Nn Mm Nn Mm Ll Kk Ll Kk Jj Ii Jj Ii Hh Gg Hh Gg Ff Ee Ff Ee Dd Cc Dd Cc Bb Aa Bb Aa|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+#else
+    const __m512i word_perm = _mm512_set_epi16(
+        19, 19, 18, 18, 17, 17, 16, 16, 15, 15, 14, 14, 13, 13, 12, 12,
+         9,  9,  8,  8,  7,  7,  6,  6,  5,  5,  4,  4,  3,  3,  2,  2
+    );
+    const __m512i permuted = _mm512_permutexvar_epi16( word_perm, blocks );
+    // This is the fallback path for CPUs that don't support VPERMB. Since we permute 16-bit groups only,
+    // VPERMB can be replaced with VPERMW. We could always use VPERMW, but at least on Tiger Lake and
+    // Ice Lake VPERMW followed by a right shift is quite noticeably slower than VPERMB.
+#endif
+
+    // Shift every odd-numbered 16-bit group to the right by 4 bits.
+    const __mmask32 shift_mask = 0xaaaaaaaa;
+    const __m512i shifted = _mm512_mask_srai_epi16( permuted, shift_mask, permuted, 4 );
+    // After applying VPSRAW, `shifted` looks like this (the "empty" nibbles are filled with zeroes):
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // | : .= :. =_  ( )[ () []  < >( <> ()  Z zY Zz Yy  X xW Xx Ww  V vU Vv Uu  T tS Tt Ss  R rQ Rr Qq
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // | P pO Pp Oo  N nM Nn Mm  L lK Ll Kk  J jI Jj Ii  H hG Hh Gg  F fE Ff Ee  D dC Dd Cc  B bA Bb Aa|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+
+    // Now we just need to zero out the higher nibble in each byte, and we're done.
+    const __m512i low_nibble_mask = _mm512_set1_epi8( 0xf );
+    return _mm512_and_si512( low_nibble_mask, shifted );
+    // The final result looks like this:
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // | :  =  .  _  (  [  )  ]  <  (  >  )  Z  Y  z  y  X  W  x  w  V  U  v  u  T  S  t  s  R  Q  r  q|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+    // | P  O  p  o  N  M  n  m  L  K  l  k  J  I  j  i  H  G  h  g  F  E  f  e  D  C  d  c  B  A  b  a|
+    // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+
+}
+
+static inline __m512 dot_q4_0_twoblocks_avx512(
     __m512 acc,
     const block_q4_0 * restrict x,
     const block_q4_0 * restrict y,
     int i
 ) {
-    // Compute combined scale for the block
-    __m512 d = _mm512_set1_ps( x[i].d * y[i].d );
+    // A pair of Q4_0 blocks spans 40 bytes, while an AVX-512 register has 64. The remaining 24 bytes
+    // can potentially be unaddressable, so we make sure to mask them out before the load, even though
+    // we don't use them at all. This might hurt the performance slightly, since the compiler is forced
+    // to use e.g. `VMOVDQU64 REG, MASK, [ADDR] + VPERMB ..., REG` instead of just `VPERMB ..., [ADDR]`.
+    const __mmask8 load_mask = 0x1f;
+    const __m512i blocks_0 = _mm512_maskz_loadu_epi64( load_mask, &x[i] );
+    const __m512i blocks_1 = _mm512_maskz_loadu_epi64( load_mask, &y[i] );
 
-    __m256i bx = bytesFromNibbles( x[i].qs );
-    __m256i by = bytesFromNibbles( y[i].qs );
+    // We want to multiply the scales, so we interpret both registers as 16 32-bit floats:
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // | 15 | 14 | 13 | 12 | 11 | 10 | 09 | 08 | 07 | 06 | 05 | 04 | 03 | 02 | 01 | 00 |
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // blocks_0_float
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // |    |    |    |    |    |    | xx | xx | xx | xx |  B | xx | xx | xx | xx |  A |
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // blocks_1_float
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // |    |    |    |    |    |    | xx | xx | xx | xx |  D | xx | xx | xx | xx |  C |
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    const __m512 blocks_0_float = _mm512_castsi512_ps( blocks_0 );
+    const __m512 blocks_1_float = _mm512_castsi512_ps( blocks_1 );
+    // We absolutely shouldn't touch the floats marked with `xx`: they contain some
+    // random data, which might very well underflow. At least on Intel, this leads
+    // to a huge penalty that can't be ignored (easily 100x or more) unless you
+    // compile your code with something like `-ffast-math` to enable FTZ/DAZ flags.
+    // (and ggml can't assume that you do)...
+    const __mmask16 scale_mul_mask = 0x21;
+#ifdef __clang__
+    // ...however, clang decides to optimize the multiplication mask away:
+    // https://godbolt.org/z/P8PqdsfvW
+    // gcc and MSVC do the sane thing. This horrible workaround forces clang to emit the mask.
+    __m512i scales;
+    __asm__(
+        "vmulps %1, %2, %0%{%3%}"
+        : "=v" ( scales )
+        : "vm" ( blocks_0_float ), "v" ( blocks_1_float ), "Yk" ( scale_mul_mask )
+    );
+#else
+    const __m512 scales = _mm512_maskz_mul_ps( scale_mul_mask, blocks_0_float, blocks_1_float );
+#endif
+    const __m512i scale_perm = _mm512_set_epi32(
+        5, 5, 5, 5, 5, 5, 5, 5,
+        0, 0, 0, 0, 0, 0, 0, 0
+    );
+    const __m512 permuted_scales = _mm512_permutexvar_ps( scale_perm, scales );
+    // After VMULPS and VPERMPS, `permuted_scales` looks like this:
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // | 15 | 14 | 13 | 12 | 11 | 10 | 09 | 08 | 07 | 06 | 05 | 04 | 03 | 02 | 01 | 00 |
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
+    // | B*D| B*D| B*D| B*D| B*D| B*D| B*D| B*D| A*C| A*C| A*C| A*C| A*C| A*C| A*C| A*C|
+    // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
 
-    // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-    const __m256i off = _mm256_set1_epi8( 8 );
-    bx = _mm256_sub_epi8( bx, off );
-    by = _mm256_sub_epi8( by, off );
+    const __m512i bytes_0 = bytes_from_q4_0_twoblocks_avx512( blocks_0 );
+    const __m512i bytes_1 = bytes_from_q4_0_twoblocks_avx512( blocks_1 );
 
-    // Sign-extend 16 signed bytes into int16_t
-    __m512i x32 = _mm512_cvtepi8_epi16( bx );
-    __m512i y32 = _mm512_cvtepi8_epi16( by );
-    // Compute products of int16_t integers, add pairwise
-    __m512i i64 = _mm512_madd_epi16( x32, y32 );
+    // Now we want to compute dot products of 4-element byte vectors and store them in
+    // 32-bit integers. That is (only one 4-element vector is shown for clarity):
+    //     +----+----+----+----+
+    // ... | 03 | 02 | 01 | 00 |
+    //     +----+----+----+----+
+    // bytes_0
+    //     +----+----+----+----+
+    // ... |  D |  C |  B |  A |
+    //     +----+----+----+----+
+    // bytes_1
+    //     +----+----+----+----+
+    // ... |  H |  G |  F |  E |
+    //     +----+----+----+----+
+    // final_res_int
+    //     +----+----+----+----+
+    // ... |  A*E+B*F+C*G+D*H  |
+    //     +----+----+----+----+
+    const __m512i plus_8 = _mm512_set1_epi8( 8 );
+    const __m512i bytes_1_minus_8 = _mm512_sub_epi8( bytes_1, plus_8 );
 
-    // Convert int32_t to float
-    __m512 p = _mm512_cvtepi32_ps( i64 );
-    // Apply the scale, and accumulate
-    return _mm512_fmadd_ps( d, p, acc );
+#ifdef __AVX512VNNI__
+    // We have VPDPBUSDS in AVX512-VNNI, which does exactly what we want, but with a catch:
+    // the *left* operand is supposed to be unsigned, while Q4_0 quantization subtracts 8
+    // from each nibble, so they can be negative. So, instead of `(bytes_0 - 8) * (bytes_1 - 8)`,
+    // we compute `bytes_0 * (bytes_1 - 8) + bytes_1 * (-8) + 64`. VPDPBUSDS uses an accumulator,
+    // which means we only need 2 instructions.
+    const __m512i dot_init = _mm512_set1_epi32( 4 * 64 );
+    const __m512i minus_8 = _mm512_set1_epi8( -8 );
+    const __m512i prod_0 = _mm512_dpbusds_epi32( dot_init, bytes_1, minus_8 );
+    const __m512i final_res_int = _mm512_dpbusds_epi32( prod_0, bytes_0, bytes_1_minus_8 );
+#else
+    // As a fallback, we have VPMADDUBSW in AVX512-BW, which uses 16-bit products instead of 32-bit ones.
+    // It has the same catch as VPDPBUSDS: the left operand should be unsigned.
+    // This is essentially the AVX-512 version of the AVX-2 trick used by GH user Const-me
+    //   ref: https://gist.github.com/Const-me/4d30e1fc767ab314596e16e90f53b6f4#file-matmultest-cpp-L119
+    const __m512i one = _mm512_set1_epi16( 1 );
+    const __m512i prod_0 = _mm512_maddubs_epi16( bytes_0, bytes_1_minus_8 );
+    const __m512i prod_1 = _mm512_maddubs_epi16( plus_8, bytes_1_minus_8 );
+    const __m512i diff = _mm512_sub_epi16( prod_0, prod_1 );
+    const __m512i final_res_int = _mm512_madd_epi16( diff, one );
+#endif
+
+    // Finally, we multiply the permuted scales and the 32-bit dot products, then accumulate.
+    const __m512 final_res_float = _mm512_cvtepi32_ps( final_res_int );
+    return _mm512_fmadd_ps( permuted_scales, final_res_float, acc );
 }
 #endif
 
@@ -2135,25 +2317,26 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
     __m512 acc0 = _mm512_setzero_ps();
     __m512 acc1 = _mm512_setzero_ps();
 
-    const int superblock_size = 8;
+    const int superblock_size = 16;
+
     const int superblock_count = nb / superblock_size;
 
     for (int superblock_ix = 0; superblock_ix < superblock_count; superblock_ix += 1) {
         int i = superblock_ix * superblock_size;
 
-        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+0 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+1 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+2 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+3 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+4 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+5 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+6 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+7 );
+        acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+0 );
+        acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+2 );
+        acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+4 );
+        acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+6 );
+        acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+8 );
+        acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+10 );
+        acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+12 );
+        acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+14 );
     }
 
     // Remainders
-    for (int i = superblock_count * superblock_size; i < nb; ++i) {
-        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i );
+    for (int i = superblock_count * superblock_size; i < nb; i += 2) {
+        acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i );
     }
 
     // Horizontal sum of all lanes of the accumulator
@@ -2373,11 +2556,11 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
             const uint8_t v0 = p0[j];
             const uint8_t v1 = p1[j];
 
-            const int8_t i0 = (int8_t) (v0 & 0xf) - 8;
-            const int8_t i1 = (int8_t) (v0 >> 4)  - 8;
+            const int i0 = (v0 & 0xf) - 8;
+            const int i1 = (v0 >> 4)  - 8;
 
-            const int8_t i2 = (int8_t) (v1 & 0xf) - 8;
-            const int8_t i3 = (int8_t) (v1 >> 4)  - 8;
+            const int i2 = (v1 & 0xf) - 8;
+            const int i3 = (v1 >> 4)  - 8;
 
             sumi += i0*i2 + i1*i3;
         }
@@ -2583,8 +2766,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     float sumf = 0.0;
 
 #if defined(__ARM_NEON)
-    float sum0 = 0.0f;
-    float sum1 = 0.0f;
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
     for (int i = 0; i < nb; i += 2) {
         const block_q4_0 * restrict x0 = &x[i + 0];
@@ -2624,14 +2807,11 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        int32x4_t p_0 = vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls);
-        int32x4_t p_1 = vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
 
-        p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs);
-        p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs);
-
-        sum0 += x0->d*y0->d*vaddvq_s32(p_0);
-        sum1 += x1->d*y1->d*vaddvq_s32(p_1);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
@@ -2643,21 +2823,17 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
         const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
 
-        const int16x8_t pl_0 = vaddq_s16(pl0l, pl0h);
-        const int16x8_t ph_0 = vaddq_s16(ph0l, ph0h);
+        const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
+        const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
+        const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
+        const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        const int16x8_t pl_1 = vaddq_s16(pl1l, pl1h);
-        const int16x8_t ph_1 = vaddq_s16(ph1l, ph1h);
-
-        const int16x8_t p_0 = vaddq_s16(pl_0, ph_0);
-        const int16x8_t p_1 = vaddq_s16(pl_1, ph_1);
-
-        sum0 += x0->d*y0->d*vaddvq_s16(p_0);
-        sum1 += x1->d*y1->d*vaddvq_s16(p_1);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
 #endif
     }
 
-    sumf = sum0 + sum1;
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -5433,6 +5609,26 @@ static void ggml_compute_forward_dup_f16(
                         }
                     }
                 }
+            } else if (dst->type == GGML_TYPE_Q4_0 || dst->type == GGML_TYPE_Q4_1) {
+                quantize_row_q_t const quantize_row_q = quantize_fns[dst->type].quantize_row_q;
+                size_t id = 0;
+                uint8_t * dst_ptr = (uint8_t *) dst->data;
+                size_t dst_row_size = nb0 * (ne00 / GGML_BLCK_SIZE[dst->type]);
+                float * src0_f32 = (float *) params->wdata;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                            // convert to f32 and quantize
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]);
+                            }
+                            quantize_row_q(src0_f32, dst_ptr + id, ne00);
+                            id += dst_row_size;
+                        }
+                    }
+                }
             } else {
                 GGML_ASSERT(false); // TODO: implement
             }
@@ -5625,6 +5821,21 @@ static void ggml_compute_forward_dup_f32(
                         }
                     }
                 }
+            } else if (dst->type == GGML_TYPE_Q4_0 || dst->type == GGML_TYPE_Q4_1) {
+                quantize_row_q_t const quantize_row_q = quantize_fns[dst->type].quantize_row_q;
+                size_t id = 0;
+                uint8_t * dst_ptr = (uint8_t *) dst->data;
+                size_t dst_row_size = nb0 * (ne00 / GGML_BLCK_SIZE[dst->type]);
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                            quantize_row_q(src0_ptr, dst_ptr + id, ne00);
+                            id += dst_row_size;
+                        }
+                    }
+                }
             } else {
                 GGML_ASSERT(false); // TODO: implement
             }
@@ -5813,6 +6024,212 @@ static void ggml_compute_forward_add_f32(
     }
 }
 
+static void ggml_compute_forward_add_f16_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    const size_t nb00 = src0->nb[0];
+    const size_t nb01 = src0->nb[1];
+
+    const size_t nb10 = src1->nb[0];
+    const size_t nb11 = src1->nb[1];
+
+    const size_t nb0 = dst->nb[0];
+    const size_t nb1 = dst->nb[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    if (nb10 == sizeof(float)) {
+        for (int j = ith; j < n; j += nth) {
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + j*nb1);
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01);
+            for (int i = 0; i < nc; i++) {
+                float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10);
+                dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + *src1_ptr);
+            }
+        }
+    }
+    else {
+        // src1 is not contiguous
+        GGML_ASSERT(false);
+    }
+}
+
+static void ggml_compute_forward_add_f16_f16(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    const size_t nb00 = src0->nb[0];
+    const size_t nb01 = src0->nb[1];
+
+    const size_t nb10 = src1->nb[0];
+    const size_t nb11 = src1->nb[1];
+
+    const size_t nb0 = dst->nb[0];
+    const size_t nb1 = dst->nb[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    if (nb10 == sizeof(ggml_fp16_t)) {
+        for (int j = ith; j < n; j += nth) {
+            ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + j*nb1);
+            ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01);
+            for (int i = 0; i < nc; i++) {
+                ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + j*nb11 + i*nb10);
+                dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(*src1_ptr));
+            }
+        }
+    }
+    else {
+        // src1 is not contiguous
+        GGML_ASSERT(false);
+    }
+}
+
+static void ggml_compute_forward_add_q_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    //const int64_t ne10 = src1->ne[0];
+    //const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+
+    //const int64_t ne0  = dst->ne[0];
+    //const int64_t ne1  = dst->ne[1];
+    const int64_t ne2  = dst->ne[2];
+    const int64_t ne3  = dst->ne[3];
+
+    const int nb00 = src0->nb[0];
+    const int nb01 = src0->nb[1];
+    const int nb02 = src0->nb[2];
+    const int nb03 = src0->nb[3];
+
+    const int nb10 = src1->nb[0];
+    const int nb11 = src1->nb[1];
+    const int nb12 = src1->nb[2];
+    const int nb13 = src1->nb[3];
+
+    const int nb0  = dst->nb[0];
+    const int nb1  = dst->nb[1];
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne03 == ne13);
+    GGML_ASSERT(ne2  == ne12);
+    GGML_ASSERT(ne3  == ne13);
+
+    const enum ggml_type type = src0->type;
+    dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
+    quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q;
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]);
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1);
+    GGML_ASSERT(dst->type == src0->type);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    // total rows in src0
+    const int nr = ne01*ne02*ne03;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    float * wdata = (float*) params->wdata + ne00 * ith;
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 indices
+        const int i03 = ir/(ne02*ne01);
+        const int i02 = (ir - i03*ne02*ne01)/ne01;
+        const int i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        // src1 and dst are same shape as src0 => same indices
+        const int i13 = i03;
+        const int i12 = i02;
+        const int i11 = i01;
+
+        const int i3 = i03;
+        const int i2 = i02;
+        const int i1 = i01;
+
+        void  * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03));
+        float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13));
+        void  * dst_row  = (void *) ((char *)  dst->data + ( i1*nb1  +  i2*nb2  +  i3*nb0));
+
+        assert(ne00 % 32 == 0);
+
+        // unquantize row from src0 to temp buffer
+        dequantize_row_q(src0_row, wdata, ne00);
+        // add src1
+        ggml_vec_acc_f32(ne00, wdata, src1_row);
+        // quantize row to dst
+        quantize_row_q(wdata, dst_row, ne00);
+    }
+}
+
 static void ggml_compute_forward_add(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -5823,6 +6240,23 @@ static void ggml_compute_forward_add(
             {
                 ggml_compute_forward_add_f32(params, src0, src1, dst);
             } break;
+        case GGML_TYPE_F16:
+            {
+                if (src1->type == GGML_TYPE_F16) {
+                    ggml_compute_forward_add_f16_f16(params, src0, src1, dst);
+                }
+                else if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add_f16_f32(params, src0, src1, dst);
+                }
+                else {
+                    GGML_ASSERT(false);
+                }
+            } break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            {
+                ggml_compute_forward_add_q_f32(params, src0, src1, dst);
+            } break;
         default:
             {
                 GGML_ASSERT(false);
@@ -7102,30 +7536,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     //}
 }
 
-static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
-    [GGML_TYPE_Q4_0] = {
-        .dequantize_row_q         = dequantize_row_q4_0,
-        .quantize_row_q           = quantize_row_q4_0,
-        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference,
-        .quantize_row_q_dot       = quantize_row_q8_0,
-        .vec_dot_q                = ggml_vec_dot_q4_0_q8_0,
-    },
-    [GGML_TYPE_Q4_1] = {
-        .dequantize_row_q         = dequantize_row_q4_1,
-        .quantize_row_q           = quantize_row_q4_1,
-        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference,
-        .quantize_row_q_dot       = quantize_row_q4_1,
-        .vec_dot_q                = ggml_vec_dot_q4_1,
-    },
-    // TODO: GGML_TYPE_Q8_0
-};
-
-// For internal test use
-quantize_fns_t ggml_internal_get_quantize_fn(size_t i) {
-    GGML_ASSERT(i < GGML_TYPE_COUNT);
-    return quantize_fns[i];
-}
-
 static void ggml_compute_forward_mul_mat_q_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -7902,11 +8312,11 @@ static void ggml_compute_forward_rope_f16(
                     const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
                           ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
-                    const float x0 = ggml_fp16_to_fp32(src[0]);
-                    const float x1 = ggml_fp16_to_fp32(src[1]);
+                    const float x0 = GGML_FP16_TO_FP32(src[0]);
+                    const float x1 = GGML_FP16_TO_FP32(src[1]);
 
-                    dst_data[0] = ggml_fp32_to_fp16(x0*cos_theta - x1*sin_theta);
-                    dst_data[1] = ggml_fp32_to_fp16(x0*sin_theta + x1*cos_theta);
+                    dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                    dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                 }
             }
         }
@@ -9982,13 +10392,29 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
             struct ggml_tensor * node = cgraph->nodes[i];
 
             switch (node->op) {
+                case GGML_OP_CPY:
                 case GGML_OP_DUP:
                     {
                         node->n_tasks = 1;
+
+                        size_t cur = 0;
+                        if (node->type == GGML_TYPE_Q4_0 || node->type == GGML_TYPE_Q4_1) {
+                            cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->ne[0];
+                        }
+
+                        work_size = MAX(work_size, cur);
                     } break;
                 case GGML_OP_ADD:
                     {
                         node->n_tasks = n_threads;
+
+                        size_t cur = 0;
+
+                        if (node->src0->type == GGML_TYPE_Q4_0 || node->src0->type == GGML_TYPE_Q4_1) {
+                            cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src0->ne[0] * n_threads;
+                        }
+
+                        work_size = MAX(work_size, cur);
                     } break;
                 case GGML_OP_SUB:
                 case GGML_OP_MUL:
@@ -10069,7 +10495,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                     {
                         node->n_tasks = n_threads;
                     } break;
-                case GGML_OP_CPY:
                 case GGML_OP_CONT:
                 case GGML_OP_RESHAPE:
                 case GGML_OP_VIEW:
@@ -11303,6 +11728,22 @@ int ggml_cpu_has_avx512(void) {
 #endif
 }
 
+int ggml_cpu_has_avx512_vbmi(void) {
+#if defined(__AVX512VBMI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_vnni(void) {
+#if defined(__AVX512VNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_fma(void) {
 #if defined(__FMA__)
     return 1;

ggml.h

@@ -430,6 +430,12 @@ struct ggml_tensor * ggml_add(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b);
 
+
+struct ggml_tensor * ggml_add_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
 struct ggml_tensor * ggml_sub(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -808,6 +814,8 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t *
 int ggml_cpu_has_avx(void);
 int ggml_cpu_has_avx2(void);
 int ggml_cpu_has_avx512(void);
+int ggml_cpu_has_avx512_vbmi(void);
+int ggml_cpu_has_avx512_vnni(void);
 int ggml_cpu_has_fma(void);
 int ggml_cpu_has_neon(void);
 int ggml_cpu_has_arm_fma(void);

llama.cpp

@@ -1,6 +1,8 @@
 // Defines fileno on msys:
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#include <cstdint>
+#include <cstdio>
 #endif
 
 #include "llama_util.h"
@@ -9,6 +11,7 @@
 #include "ggml.h"
 
 #include <array>
+#include <ctime>
 #include <cinttypes>
 #include <fstream>
 #include <random>
@@ -41,35 +44,51 @@ static const size_t MB = 1024*1024;
 // TODO: dynamically determine these sizes
 //       needs modifications in ggml
 
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
-    { MODEL_7B,    512ull*MB },
-    { MODEL_13B,   512ull*MB },
-    { MODEL_30B,   512ull*MB },
-    { MODEL_65B,   512ull*MB },
-};
+static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
+{
+    static std::map<e_model, size_t> _MEM_REQ_SCRATCH0 = {
+        { MODEL_7B,    512ull * MB },
+        { MODEL_13B,   512ull * MB },
+        { MODEL_30B,   512ull * MB },
+        { MODEL_65B,   512ull * MB },
+    };
+    return _MEM_REQ_SCRATCH0;
+}
 
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
-    { MODEL_7B,    512ull*MB },
-    { MODEL_13B,   512ull*MB },
-    { MODEL_30B,   512ull*MB },
-    { MODEL_65B,   512ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
+{
+    static std::map<e_model, size_t> _MEM_REQ_SCRATCH1 = {
+        { MODEL_7B,    512ull * MB },
+        { MODEL_13B,   512ull * MB },
+        { MODEL_30B,   512ull * MB },
+        { MODEL_65B,   512ull * MB },
+    };
+    return _MEM_REQ_SCRATCH1;
 };
 
 // 2*n_embd*n_ctx*n_layer*sizeof(float16)
-static const std::map<e_model, size_t> MEM_REQ_KV_SELF = {
-    { MODEL_7B,   1026ull*MB },
-    { MODEL_13B,  1608ull*MB },
-    { MODEL_30B,  3124ull*MB },
-    { MODEL_65B,  5120ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
+{
+    static std::map<e_model, size_t> _MEM_REQ_KV_SELF = {
+        { MODEL_7B,   1026ull * MB },
+        { MODEL_13B,  1608ull * MB },
+        { MODEL_30B,  3124ull * MB },
+        { MODEL_65B,  5120ull * MB },
+    };
+    return _MEM_REQ_KV_SELF;
 };
 
 // this is mostly needed for temporary mul_mat buffers to dequantize the data
 // not actually needed if BLAS is disabled
-static const std::map<e_model, size_t> MEM_REQ_EVAL = {
-    { MODEL_7B,   768ull*MB },
-    { MODEL_13B, 1024ull*MB },
-    { MODEL_30B, 1280ull*MB },
-    { MODEL_65B, 1536ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_EVAL()
+{
+    static std::map<e_model, size_t> _MEM_REQ_EVAL = {
+        { MODEL_7B,   768ull * MB },
+        { MODEL_13B, 1024ull * MB },
+        { MODEL_30B, 1280ull * MB },
+        { MODEL_65B, 1536ull * MB },
+    };
+    return _MEM_REQ_EVAL;
 };
 
 // default hparams (LLaMA 7B)
@@ -261,12 +280,12 @@ static size_t checked_div(size_t a, size_t b) {
 }
 
 static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
-    std::string ret = "[" + std::to_string(ne.at(0));
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5u", ne.at(0));
     for (size_t i = 1; i < ne.size(); i++) {
-        ret += " x " + std::to_string(ne.at(i));
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), " x %5u", ne.at(i));
     }
-    ret += "]";
-    return ret;
+    return buf;
 }
 
 static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml_type type) {
@@ -616,6 +635,7 @@ struct llama_model_loader {
             throw format("llama.cpp: tensor '%s' has wrong shape; expected %s, got %s",
                          name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str());
         }
+
         return get_tensor_for(lt);
     }
 
@@ -898,13 +918,13 @@ static void llama_model_load_internal(
         const size_t mem_required =
             ctx_size +
             mmapped_size +
-            MEM_REQ_SCRATCH0.at(model.type) +
-            MEM_REQ_SCRATCH1.at(model.type) +
-            MEM_REQ_EVAL.at    (model.type);
+            MEM_REQ_SCRATCH0().at(model.type) +
+            MEM_REQ_SCRATCH1().at(model.type) +
+            MEM_REQ_EVAL().at(model.type);
 
         // this is the memory required by one llama_state
         const size_t mem_required_state =
-            scale*MEM_REQ_KV_SELF.at(model.type);
+            scale*MEM_REQ_KV_SELF().at(model.type);
 
         fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
@@ -941,8 +961,8 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab});
-        model.norm   = ml->get_tensor("norm.weight", {n_embd});
-        model.output = ml->get_tensor("output.weight", {n_embd, n_vocab});
+        model.norm           = ml->get_tensor("norm.weight",           {n_embd});
+        model.output         = ml->get_tensor("output.weight",         {n_embd, n_vocab});
 
         model.layers.resize(n_layer);
         for (uint32_t i = 0; i < n_layer; ++i) {
@@ -1569,7 +1589,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         tensor.data = read_data.addr;
         model_loader->load_data_for(tensor);
 
-        printf("[%zu/%zu] %36s - %s, type = %6s, ",
+        printf("[%4zu/%4zu] %36s - %16s, type = %6s, ",
                ++idx, model_loader->tensors_map.tensors.size(),
                tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
                ggml_type_name(tensor.type));
@@ -1731,10 +1751,10 @@ struct llama_context * llama_init_from_file(
             ctx->embedding.resize(hparams.n_embd);
         }
 
-        ctx->buf_compute.resize(MEM_REQ_EVAL.at(ctx->model.type));
+        ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type));
 
-        ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0.at(ctx->model.type));
-        ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1.at(ctx->model.type));
+        ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0().at(ctx->model.type));
+        ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type));
     }
 
     return ctx;
@@ -1757,6 +1777,254 @@ int llama_model_quantize(
     }
 }
 
+int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+    fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
+
+    auto & model = ctx->model;
+
+    const int64_t t_start_lora_us = ggml_time_us();
+
+    auto fin = std::ifstream(path_lora, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora);
+        return 1;
+    }
+
+    // verify magic and version
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != 'ggla') {
+            fprintf(stderr, "%s: bad file magic\n", __func__);
+            return 1;
+        }
+        uint32_t format_version;
+        fin.read((char *) &format_version, sizeof(format_version));
+
+        if (format_version != 1) {
+            fprintf(stderr, "%s: unsupported file version\n", __func__ );
+            return 1;
+        }
+    }
+
+    int32_t lora_r;
+    int32_t lora_alpha;
+    fin.read((char *) &lora_r, sizeof(lora_r));
+    fin.read((char *) &lora_alpha, sizeof(lora_alpha));
+    float scaling = (float)lora_alpha / (float)lora_r;
+
+    fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
+
+
+    // create a temporary ggml context to store the lora tensors
+    // todo: calculate size from biggest possible tensor
+    std::vector<uint8_t> lora_buf(1024ull * 1024ull * 1024ull);
+    struct ggml_init_params params;
+    params.mem_size   = lora_buf.size();
+    params.mem_buffer = lora_buf.data();
+    params.no_alloc   = false;
+
+    ggml_context * lora_ctx = ggml_init(params);
+    std::unordered_map<std::string, struct ggml_tensor *> lora_tensors;
+
+    // create a name -> tensor map of the model to accelerate lookups
+    std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
+    for (auto & kv: model.tensors_by_name) {
+        model_tensors.insert(kv);
+    }
+
+
+    // load base model
+    std::unique_ptr<llama_model_loader> model_loader;
+    ggml_context * base_ctx = NULL;
+    llama_buffer base_buf;
+    if (path_base_model) {
+        fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model);
+        model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*vocab_only*/ false));
+
+        size_t ctx_size, mmapped_size;
+        model_loader->calc_sizes(&ctx_size, &mmapped_size);
+        base_buf.resize(ctx_size);
+
+        ggml_init_params base_params;
+        base_params.mem_size   = base_buf.size;
+        base_params.mem_buffer = base_buf.addr;
+        base_params.no_alloc   = model_loader->use_mmap;
+
+        base_ctx = ggml_init(base_params);
+
+        model_loader->ggml_ctx = base_ctx;
+
+        // maybe this should in llama_model_loader
+        if (model_loader->use_mmap) {
+            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ false));
+        }
+    }
+
+    // read tensors and apply
+    bool warned = false;
+    int n_tensors = 0;
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ftype;
+
+        fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+        fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+        if (fin.eof()) {
+            break;
+        }
+
+        int32_t ne[2] = { 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+
+        std::string name(length, 0);
+        fin.read(&name[0], length);
+
+        // check for lora suffix and get the type of tensor
+        const std::string lora_suffix = ".lora";
+        size_t pos = name.rfind(lora_suffix);
+        if (pos == std::string::npos) {
+            fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
+            return 1;
+        }
+
+        std::string lora_type = name.substr(pos + lora_suffix.length());
+        std::string base_name = name;
+        base_name.erase(pos);
+        // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
+
+        if (model_tensors.find(base_name.data()) == model_tensors.end()) {
+            fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
+            return 1;
+        }
+
+        // create ggml tensor
+        ggml_type wtype;
+        switch (ftype) {
+            case 0: wtype = GGML_TYPE_F32;  break;
+            case 1: wtype = GGML_TYPE_F16;  break;
+            default:
+                    {
+                        fprintf(stderr, "%s: invalid tensor data type '%d'\n",
+                                __func__, ftype);
+                        return false;
+                    }
+        }
+        ggml_tensor* lora_tensor;
+        if (n_dims == 2) {
+            lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
+        }
+        else {
+            fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims);
+            return 1;
+        }
+
+        // load tensor data
+        size_t offset = fin.tellg();
+        size_t tensor_data_size = ggml_nbytes(lora_tensor);
+        offset = (offset + 31) & -32;
+        fin.seekg(offset);
+        fin.read((char*)lora_tensor->data, tensor_data_size);
+
+        lora_tensors[name] = lora_tensor;
+
+        // check if we have both A and B tensors and apply
+        if (lora_tensors.find(base_name + ".loraA") != lora_tensors.end() &&
+            lora_tensors.find(base_name + ".loraB") != lora_tensors.end()) {
+
+            ggml_tensor * dest_t = model_tensors[base_name];
+            ggml_tensor * base_t;
+            if (model_loader) {
+                // load from base model
+                if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) {
+                    fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
+                    return 1;
+                }
+                size_t idx = model_loader->tensors_map.name_to_idx[base_name];
+                llama_load_tensor & lt = model_loader->tensors_map.tensors[idx];
+                base_t = model_loader->get_tensor(base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] });
+                lt.data = (uint8_t *) lt.ggml_tensor->data;
+                model_loader->load_data_for(lt);
+                lt.ggml_tensor->data = lt.data;
+            }
+            else {
+                base_t = dest_t;
+            }
+
+            if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1) {
+                if (!warned) {
+                    fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
+                                    "use a f16 or f32 base model with --lora-base\n", __func__);
+                    warned = true;
+                }
+            }
+
+            ggml_tensor * loraA = lora_tensors[base_name + ".loraA"];
+            ggml_tensor * loraB = lora_tensors[base_name + ".loraB"];
+
+            if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
+                fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
+                               " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
+                return 1;
+            }
+
+            // w = w + BA*s
+            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+
+            if (scaling != 1.0f) {
+                ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
+                BA = ggml_scale(lora_ctx, BA, scale_tensor);
+            }
+
+            ggml_tensor * r;
+            if (base_t == dest_t) {
+                r = ggml_add_inplace(lora_ctx, dest_t, BA);
+            }
+            else {
+                r = ggml_add(lora_ctx, base_t, BA);
+                r = ggml_cpy(lora_ctx, r, dest_t);
+            }
+
+            struct ggml_cgraph gf = ggml_build_forward(r);
+            gf.n_threads = n_threads;
+            ggml_graph_compute(lora_ctx, &gf);
+
+            // we won't need these tensors again, reset the context to save memory
+            ggml_free(lora_ctx);
+            lora_ctx = ggml_init(params);
+            lora_tensors.clear();
+
+            n_tensors++;
+            if (n_tensors % 4 == 0)
+                fprintf(stderr, ".");
+        }
+    }
+
+    // TODO: this should be in a destructor, it will leak on failure
+    ggml_free(lora_ctx);
+    if (base_ctx) {
+        ggml_free(base_ctx);
+    }
+
+    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
+    fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0);
+
+    return 0;
+}
+
+int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+    try {
+        return llama_apply_lora_from_file_internal(ctx, path_lora, path_base_model, n_threads);
+    } catch (const std::string & err) {
+        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.c_str());
+        return 1;
+    }
+}
+
 // Returns the KV cache that will contain the context for the
 // ongoing prediction with the model.
 const uint8_t * llama_get_kv_cache(struct llama_context * ctx) {
@@ -1914,18 +2182,20 @@ const char * llama_print_system_info(void) {
     static std::string s;
 
     s  = "";
-    s += "AVX = "       + std::to_string(ggml_cpu_has_avx())       + " | ";
-    s += "AVX2 = "      + std::to_string(ggml_cpu_has_avx2())      + " | ";
-    s += "AVX512 = "    + std::to_string(ggml_cpu_has_avx512())    + " | ";
-    s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
-    s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
-    s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
-    s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
-    s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
-    s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
-    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
-    s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
-    s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
+    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
+    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
+    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
+    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
+    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
+    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
+    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
+    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
+    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
+    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
+    s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
+    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
+    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
 
     return s.c_str();
 }

llama.h

@@ -96,6 +96,18 @@ extern "C" {
             const char * fname_out,
       enum llama_ftype   ftype);
 
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
+
     // Returns the KV cache that will contain the context for the
     // ongoing prediction with the model.
     LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);

llama_util.h

@@ -43,8 +43,12 @@
     } while (0)
 
 #ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
 __attribute__((format(printf, 1, 2)))
 #endif
+#endif
 static std::string format(const char * fmt, ...) {
     va_list ap, ap2;
     va_start(ap, fmt);
@@ -57,7 +61,7 @@ static std::string format(const char * fmt, ...) {
     va_end(ap2);
     va_end(ap);
     return std::string(buf.data(), size);
-};
+}
 
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
@@ -164,7 +168,7 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
@@ -172,15 +176,16 @@ struct llama_mmap {
         flags |= MAP_POPULATE;
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        close(fd);
         if (addr == MAP_FAILED) {
             throw format("mmap failed: %s", strerror(errno));
         }
 
-        // Advise the kernel to preload the mapped memory
-        if (madvise(addr, file->size, MADV_WILLNEED)) {
-            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
-                    strerror(errno));
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, file->size, MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
         }
     }
 
@@ -190,14 +195,13 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
         DWORD error = GetLastError();
-        CloseHandle(hFile);
 
         if (hMapping == NULL) {
             throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
@@ -212,13 +216,15 @@ struct llama_mmap {
         }
 
         #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
-        // Advise the kernel to preload the mapped memory
-        WIN32_MEMORY_RANGE_ENTRY range;
-        range.VirtualAddress = addr;
-        range.NumberOfBytes = (SIZE_T)size;
-        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
         }
         #else
         #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")

pocs/CMakeLists.txt

@@ -0,0 +1,12 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+if (EMSCRIPTEN)
+else()
+    add_subdirectory(vdot)
+endif()

pocs/vdot/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET vdot)
+add_executable(${TARGET} vdot.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

pocs/vdot/vdot.cpp

@@ -0,0 +1,305 @@
+#include <cstdio>
+#include <vector>
+#include <random>
+#include <chrono>
+#include <cstdlib>
+#include <cmath>
+#include <cassert>
+#include <cstring>
+#include <array>
+
+#include <ggml.h>
+
+constexpr int kVecSize = 1 << 18;
+
+float drawFromGaussianPdf(std::mt19937& rndm) {
+    constexpr double kScale = 1./(1. + std::mt19937::max());
+    constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale;
+    static float lastX;
+    static bool haveX = false;
+    if (haveX) { haveX = false; return lastX; }
+    auto r = sqrt(-2*log(1 - kScale*rndm()));
+    auto phi = kTwoPiTimesScale * rndm();
+    lastX = r*sin(phi);
+    haveX = true;
+    return r*cos(phi);
+}
+void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
+    for (auto& v : values) v = mean + drawFromGaussianPdf(rndm);
+}
+
+// Copy-pasted from ggml.c
+#define QK4_0 32
+typedef struct {
+    float   d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;          // delta
+    float   m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+// Copy-pasted from ggml.c
+#define QK8_0 32
+typedef struct {
+    float   d;          // delta
+    int8_t  qs[QK8_0];  // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+
+// "Scalar" dot product between the quantized vector x and float vector y
+inline double dot(int n, const block_q4_0* x, const float* y) {
+    const static float kValues[16] = {-8.f, -7.f, -6.f, -5.f, -4.f, -3.f, -2.f, -1.f, 0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f};
+    constexpr uint32_t kMask1 = 0x0f0f0f0f;
+    uint32_t u1, u2;
+    auto q1 = (const uint8_t*)&u1;
+    auto q2 = (const uint8_t*)&u2;
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        float d = x->d;
+        auto u = (const uint32_t*)x->qs;
+        float s = 0;
+        for (int k=0; k<4; ++k) {
+            u1 = u[k] & kMask1;
+            u2 = (u[k] >> 4) & kMask1;
+            s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] +
+                 y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] +
+                 y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] +
+                 y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]];
+            y += 8;
+        }
+        sum += s*d;
+        ++x;
+    }
+    return sum;
+}
+// Alternative version of the above. Faster on my Mac (~45 us vs ~55 us per dot product),
+// but about the same on X86_64 (Ryzen 7950X CPU).
+inline double dot3(int n, const block_q4_0* x, const float* y) {
+    const static std::pair<float,float> kValues[256] = {
+        {-8.f, -8.f}, {-7.f, -8.f}, {-6.f, -8.f}, {-5.f, -8.f}, {-4.f, -8.f}, {-3.f, -8.f}, {-2.f, -8.f}, {-1.f, -8.f},
+        { 0.f, -8.f}, { 1.f, -8.f}, { 2.f, -8.f}, { 3.f, -8.f}, { 4.f, -8.f}, { 5.f, -8.f}, { 6.f, -8.f}, { 7.f, -8.f},
+        {-8.f, -7.f}, {-7.f, -7.f}, {-6.f, -7.f}, {-5.f, -7.f}, {-4.f, -7.f}, {-3.f, -7.f}, {-2.f, -7.f}, {-1.f, -7.f},
+        { 0.f, -7.f}, { 1.f, -7.f}, { 2.f, -7.f}, { 3.f, -7.f}, { 4.f, -7.f}, { 5.f, -7.f}, { 6.f, -7.f}, { 7.f, -7.f},
+        {-8.f, -6.f}, {-7.f, -6.f}, {-6.f, -6.f}, {-5.f, -6.f}, {-4.f, -6.f}, {-3.f, -6.f}, {-2.f, -6.f}, {-1.f, -6.f},
+        { 0.f, -6.f}, { 1.f, -6.f}, { 2.f, -6.f}, { 3.f, -6.f}, { 4.f, -6.f}, { 5.f, -6.f}, { 6.f, -6.f}, { 7.f, -6.f},
+        {-8.f, -5.f}, {-7.f, -5.f}, {-6.f, -5.f}, {-5.f, -5.f}, {-4.f, -5.f}, {-3.f, -5.f}, {-2.f, -5.f}, {-1.f, -5.f},
+        { 0.f, -5.f}, { 1.f, -5.f}, { 2.f, -5.f}, { 3.f, -5.f}, { 4.f, -5.f}, { 5.f, -5.f}, { 6.f, -5.f}, { 7.f, -5.f},
+        {-8.f, -4.f}, {-7.f, -4.f}, {-6.f, -4.f}, {-5.f, -4.f}, {-4.f, -4.f}, {-3.f, -4.f}, {-2.f, -4.f}, {-1.f, -4.f},
+        { 0.f, -4.f}, { 1.f, -4.f}, { 2.f, -4.f}, { 3.f, -4.f}, { 4.f, -4.f}, { 5.f, -4.f}, { 6.f, -4.f}, { 7.f, -4.f},
+        {-8.f, -3.f}, {-7.f, -3.f}, {-6.f, -3.f}, {-5.f, -3.f}, {-4.f, -3.f}, {-3.f, -3.f}, {-2.f, -3.f}, {-1.f, -3.f},
+        { 0.f, -3.f}, { 1.f, -3.f}, { 2.f, -3.f}, { 3.f, -3.f}, { 4.f, -3.f}, { 5.f, -3.f}, { 6.f, -3.f}, { 7.f, -3.f},
+        {-8.f, -2.f}, {-7.f, -2.f}, {-6.f, -2.f}, {-5.f, -2.f}, {-4.f, -2.f}, {-3.f, -2.f}, {-2.f, -2.f}, {-1.f, -2.f},
+        { 0.f, -2.f}, { 1.f, -2.f}, { 2.f, -2.f}, { 3.f, -2.f}, { 4.f, -2.f}, { 5.f, -2.f}, { 6.f, -2.f}, { 7.f, -2.f},
+        {-8.f, -1.f}, {-7.f, -1.f}, {-6.f, -1.f}, {-5.f, -1.f}, {-4.f, -1.f}, {-3.f, -1.f}, {-2.f, -1.f}, {-1.f, -1.f},
+        { 0.f, -1.f}, { 1.f, -1.f}, { 2.f, -1.f}, { 3.f, -1.f}, { 4.f, -1.f}, { 5.f, -1.f}, { 6.f, -1.f}, { 7.f, -1.f},
+        {-8.f,  0.f}, {-7.f,  0.f}, {-6.f,  0.f}, {-5.f,  0.f}, {-4.f,  0.f}, {-3.f,  0.f}, {-2.f,  0.f}, {-1.f,  0.f},
+        { 0.f,  0.f}, { 1.f,  0.f}, { 2.f,  0.f}, { 3.f,  0.f}, { 4.f,  0.f}, { 5.f,  0.f}, { 6.f,  0.f}, { 7.f,  0.f},
+        {-8.f,  1.f}, {-7.f,  1.f}, {-6.f,  1.f}, {-5.f,  1.f}, {-4.f,  1.f}, {-3.f,  1.f}, {-2.f,  1.f}, {-1.f,  1.f},
+        { 0.f,  1.f}, { 1.f,  1.f}, { 2.f,  1.f}, { 3.f,  1.f}, { 4.f,  1.f}, { 5.f,  1.f}, { 6.f,  1.f}, { 7.f,  1.f},
+        {-8.f,  2.f}, {-7.f,  2.f}, {-6.f,  2.f}, {-5.f,  2.f}, {-4.f,  2.f}, {-3.f,  2.f}, {-2.f,  2.f}, {-1.f,  2.f},
+        { 0.f,  2.f}, { 1.f,  2.f}, { 2.f,  2.f}, { 3.f,  2.f}, { 4.f,  2.f}, { 5.f,  2.f}, { 6.f,  2.f}, { 7.f,  2.f},
+        {-8.f,  3.f}, {-7.f,  3.f}, {-6.f,  3.f}, {-5.f,  3.f}, {-4.f,  3.f}, {-3.f,  3.f}, {-2.f,  3.f}, {-1.f,  3.f},
+        { 0.f,  3.f}, { 1.f,  3.f}, { 2.f,  3.f}, { 3.f,  3.f}, { 4.f,  3.f}, { 5.f,  3.f}, { 6.f,  3.f}, { 7.f,  3.f},
+        {-8.f,  4.f}, {-7.f,  4.f}, {-6.f,  4.f}, {-5.f,  4.f}, {-4.f,  4.f}, {-3.f,  4.f}, {-2.f,  4.f}, {-1.f,  4.f},
+        { 0.f,  4.f}, { 1.f,  4.f}, { 2.f,  4.f}, { 3.f,  4.f}, { 4.f,  4.f}, { 5.f,  4.f}, { 6.f,  4.f}, { 7.f,  4.f},
+        {-8.f,  5.f}, {-7.f,  5.f}, {-6.f,  5.f}, {-5.f,  5.f}, {-4.f,  5.f}, {-3.f,  5.f}, {-2.f,  5.f}, {-1.f,  5.f},
+        { 0.f,  5.f}, { 1.f,  5.f}, { 2.f,  5.f}, { 3.f,  5.f}, { 4.f,  5.f}, { 5.f,  5.f}, { 6.f,  5.f}, { 7.f,  5.f},
+        {-8.f,  6.f}, {-7.f,  6.f}, {-6.f,  6.f}, {-5.f,  6.f}, {-4.f,  6.f}, {-3.f,  6.f}, {-2.f,  6.f}, {-1.f,  6.f},
+        { 0.f,  6.f}, { 1.f,  6.f}, { 2.f,  6.f}, { 3.f,  6.f}, { 4.f,  6.f}, { 5.f,  6.f}, { 6.f,  6.f}, { 7.f,  6.f},
+        {-8.f,  7.f}, {-7.f,  7.f}, {-6.f,  7.f}, {-5.f,  7.f}, {-4.f,  7.f}, {-3.f,  7.f}, {-2.f,  7.f}, {-1.f,  7.f},
+        { 0.f,  7.f}, { 1.f,  7.f}, { 2.f,  7.f}, { 3.f,  7.f}, { 4.f,  7.f}, { 5.f,  7.f}, { 6.f,  7.f}, { 7.f,  7.f}
+    };
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        float d = x->d;
+        auto q = x->qs;
+        float s = 0;
+        for (int k=0; k<4; ++k) {
+            s += y[0]*kValues[q[0]].first + y[1]*kValues[q[0]].second +
+                 y[2]*kValues[q[1]].first + y[3]*kValues[q[1]].second +
+                 y[4]*kValues[q[2]].first + y[5]*kValues[q[2]].second +
+                 y[6]*kValues[q[3]].first + y[7]*kValues[q[3]].second;
+            y += 8; q += 4;
+        }
+        sum += s*d;
+        ++x;
+    }
+    return sum;
+}
+
+inline double dot41(int n, const block_q4_1* x, const float* y) {
+    const static float kValues[16] = {0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f, 9.f, 10.f, 11.f, 12.f, 13.f, 14.f, 15.f};
+    constexpr uint32_t kMask1 = 0x0f0f0f0f;
+    uint32_t u1, u2;
+    auto q1 = (const uint8_t*)&u1;
+    auto q2 = (const uint8_t*)&u2;
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        auto u = (const uint32_t*)x->qs;
+        float s = 0, s1 = 0;
+        for (int k=0; k<4; ++k) {
+            u1 = u[k] & kMask1;
+            u2 = (u[k] >> 4) & kMask1;
+            s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] +
+                 y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] +
+                 y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] +
+                 y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]];
+            s1 += y[0] + y[1] + y[2] + y[3] + y[4] + y[5] + y[6] + y[7];
+            y += 8;
+        }
+        sum += s*x->d + s1*x->m;
+        ++x;
+    }
+    return sum;
+}
+
+// Copy-pasted from ggml.c
+static void quantize_row_q8_0_reference(const float *x, block_q8_0 *y, int k) {
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int l = 0; l < QK8_0; l++) {
+            const float v = x[i*QK8_0 + l];
+            amax = std::max(amax, fabsf(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+
+        for (int l = 0; l < QK8_0; ++l) {
+            const float   v  = x[i*QK8_0 + l]*id;
+            y[i].qs[l] = roundf(v);
+        }
+    }
+}
+
+// Copy-pasted from ggml.c
+static void dot_q4_q8(const int n, float* s, const void* vx, const void* vy) {
+    const int nb = n / QK8_0;
+    const block_q4_0* x = (const block_q4_0*)vx;
+    const block_q8_0* y = (const block_q8_0*)vy;
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+        const float d0 = x[i].d;
+        const float d1 = y[i].d;
+
+        const uint8_t * p0 = x[i].qs;
+        const  int8_t * p1 = y[i].qs;
+
+        int sumi = 0;
+        for (int j = 0; j < QK8_0/2; j++) {
+            const uint8_t v0 = p0[j];
+
+            const int i0 = (int8_t) (v0 & 0xf) - 8;
+            const int i1 = (int8_t) (v0 >> 4)  - 8;
+
+            const int i2 = p1[2*j + 0];
+            const int i3 = p1[2*j + 1];
+
+            sumi += i0*i2 + i1*i3;
+        }
+        sumf += d0*d1*sumi;
+    }
+    *s = sumf;
+}
+
+int main(int argc, char** argv) {
+
+    int nloop = argc > 1 ? atoi(argv[1]) : 10;
+    bool scalar = argc > 2 ? atoi(argv[2]) : false;
+    bool useQ4_1 = argc > 3 ? atoi(argv[3]) : false;
+
+    if (scalar && useQ4_1) {
+        printf("It is not possible to use Q4_1 quantization and scalar implementations\n");
+        return 1;
+    }
+
+    std::mt19937 rndm(1234);
+
+    std::vector<float> x1(kVecSize), y1(kVecSize);
+    int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
+    int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
+
+    auto funcs = useQ4_1 ? ggml_internal_get_quantize_fn(GGML_TYPE_Q4_1) : ggml_internal_get_quantize_fn(GGML_TYPE_Q4_0);
+
+    std::vector<block_q4_0> q40;
+    std::vector<block_q4_1> q41;
+    if (useQ4_1) q41.resize(n4);
+    else q40.resize(n4);
+    std::vector<block_q8_0> q8(n8);
+    std::vector<int64_t> H(16, 0);
+    double sumt = 0, sumt2 = 0, maxt = 0;
+    double sumqt = 0, sumqt2 = 0, maxqt = 0;
+    double sum = 0, sumq = 0, exactSum = 0;
+    for (int iloop=0; iloop<nloop; ++iloop) {
+
+        // Fill vector x with random numbers
+        fillRandomGaussianFloats(x1, rndm);
+
+        // Fill vector y with random numbers
+        fillRandomGaussianFloats(y1, rndm);
+
+        // Compute the exact dot product
+        for (int k=0; k<kVecSize; ++k) exactSum += x1[k]*y1[k];
+
+        // quantize x.
+        // Note, we do not include this in the timing as in practical application
+        // we already have the quantized model weights.
+        if (useQ4_1) {
+            funcs.quantize_row_q(x1.data(), q41.data(), kVecSize);
+        } else {
+            funcs.quantize_row_q(x1.data(), q40.data(), kVecSize);
+        }
+
+        // Now measure time the dot product needs using the "scalar" version above
+        auto t1 = std::chrono::high_resolution_clock::now();
+        if (useQ4_1) sum += dot41(kVecSize / QK4_1, q41.data(), y1.data());
+        else sum += dot(kVecSize / QK4_0, q40.data(), y1.data());
+        auto t2 = std::chrono::high_resolution_clock::now();
+        auto t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        sumt += t; sumt2 += t*t; maxt = std::max(maxt, t);
+
+        // And now measure the time needed to quantize y and perform the dot product with the quantized y
+        t1 = std::chrono::high_resolution_clock::now();
+        float result;
+        if (scalar) {
+            quantize_row_q8_0_reference(y1.data(), q8.data(), kVecSize);
+            dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
+        }
+        else {
+            funcs.quantize_row_q_dot(y1.data(), q8.data(), kVecSize);
+            if (useQ4_1) funcs.vec_dot_q(kVecSize, &result, q41.data(), q8.data());
+            else funcs.vec_dot_q(kVecSize, &result, q40.data(), q8.data());
+        }
+        sumq += result;
+        t2 = std::chrono::high_resolution_clock::now();
+        t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        sumqt += t; sumqt2 += t*t; maxqt = std::max(maxqt, t);
+
+    }
+
+    // Report the time (and the average of the dot products so the compiler does not come up with the idea
+    // of optimizing away the function calls after figuring that the result is not used).
+    sum /= nloop; sumq /= nloop;
+    exactSum /= nloop;
+    printf("Exact result: <dot> = %g\n",exactSum);
+    printf("<dot> = %g, %g\n",sum,sumq);
+    sumt /= nloop; sumt2 /= nloop; sumt2 -= sumt*sumt;
+    if (sumt2 > 0) sumt2 = sqrt(sumt2);
+    printf("time = %g +/- %g us. maxt = %g us\n",sumt,sumt2,maxt);
+    sumqt /= nloop; sumqt2 /= nloop; sumqt2 -= sumqt*sumqt;
+    if (sumqt2 > 0) sumqt2 = sqrt(sumqt2);
+    printf("timeq = %g +/- %g us. maxt = %g us\n",sumqt,sumqt2,maxqt);
+    return 0;
+}

tests/test-tokenizer-0.cpp

@@ -5,13 +5,17 @@
 #include <map>
 #include <vector>
 
-static const std::map<std::string, std::vector<llama_token>> k_tests = {
-    { "Hello World",        { 1,  10994,   2787, }, },
-    { " Hello World",       { 1,  15043,   2787, }, },
-    { " Hello World!",      { 1,  15043,   2787,  29991, }, },
-    { " this is 🦙.cpp",    { 1,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
-    { "w048 7tuijk dsdfhu", { 1,  29893,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
-    { "нещо на Български",  { 1,    821,   4851,    665,   1386,  29713,   1305, }, },
+static const std::map<std::string, std::vector<llama_token>> & k_tests()
+{
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { "Hello World",        { 1,  10994,   2787, }, },
+        { " Hello World",       { 1,  15043,   2787, }, },
+        { " Hello World!",      { 1,  15043,   2787,  29991, }, },
+        { " this is 🦙.cpp",    { 1,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
+        { "w048 7tuijk dsdfhu", { 1,  29893,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
+        { "нещо на Български",  { 1,    821,   4851,    665,   1386,  29713,   1305, }, },
+    };
+    return _k_tests;
 };
 
 int main(int argc, char **argv) {
@@ -47,7 +51,7 @@ int main(int argc, char **argv) {
         return 2;
     }
 
-    for (const auto & test_kv : k_tests) {
+    for (const auto & test_kv : k_tests()) {
         std::vector<llama_token> res(test_kv.first.size());
         const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
         res.resize(n);